Transformer Vibration Forecasting for Advancing Rail Safety and Maintenance 4.0

TL;DR

Railway axles experience non-stationary vibration signals due to varying speed and load, creating a need for robust predictive maintenance beyond periodic inspections. The paper introduces two Transformer-based forecasting architectures, sf and ssf, where sf uses time-domain processing with ProbSparse attention and an autoregressive decoder, and ssf extends this framework to the frequency domain with spectrogram inputs, HiLo-attention, and a two-step decoder with spectral filtering. The authors provide a comprehensive data pipeline, combining experimental bogie-rig measurements with Abaqus simulations, and demonstrate that SSF delivers superior time- and frequency-domain forecasting, along with capabilities for missing-data imputation and outlier detection. These results advance Maintenance 4.0 efforts and lay groundwork for digital-twin-style tools, such as a bogie digital twin, to enhance rail safety, reliability, and maintenance efficiency.

Abstract

Maintaining railway axles is critical to preventing severe accidents and financial losses. The railway industry is increasingly interested in advanced condition monitoring techniques to enhance safety and efficiency, moving beyond traditional periodic inspections toward Maintenance 4.0. This study introduces a robust Deep Autoregressive solution that integrates seamlessly with existing systems to avert mechanical failures. Our approach simulates and predicts vibration signals under various conditions and fault scenarios, improving dataset robustness for more effective detection systems. These systems can alert maintenance needs, preventing accidents preemptively. We use experimental vibration signals from accelerometers on train axles. Our primary contributions include a transformer model, ShaftFormer, designed for processing time series data, and an alternative model incorporating spectral methods and enhanced observation models. Simulating vibration signals under diverse conditions mitigates the high cost of obtaining experimental signals for all scenarios. Given the non-stationary nature of railway vibration signals, influenced by speed and load changes, our models address these complexities, offering a powerful tool for predictive maintenance in the rail industry.

Figures (18)

• Figure 1: Test rig with Bogie Y21 set, including tested wheelset (1), fixed wheelset (2), lhs axle box (3), rhs axle box (4), motor and driving system (5)
• Figure 2: Architecture of the sf, showing 1) elt: the preprocessing layers for the signals and their features, 2) the encoder and the generated encoder memory, 3) the decoder and the resulting prediction. The input of the encoder, called source, corresponds to historical data, while the input of the decoder, the target, corresponds to the true values of the future time steps which are going to be predicted. This method is called teacher forcing and is used when training the model. It is used widely in transformer architectures vaswani2017attention.
• Figure 3: Architecture of the sf's encoder based on the Informer, using the ProbSparse self-attention layer, in purple, and the convolutional block, in green.
• Figure 4: Output of the sf model
• Figure 5: Encoder of the ssf